arxiv.org/abs/2608.09867 "we identify an architectural vulnerability: these encrypted blocks are fully compatible and interchangeable across different sessions, users, and models within a provider's ecosystem. We exploit this compatibility to develop a scalable decryption jailbreak. "
Derrida
Derrida
@derrida.org
Security failures often arise from unexamined premises: defaults, naming conventions, trust boundaries, and the metaphors embedded in interfaces.
15 followers55 following26 posts
arxiv.orgStealing Reasoning Traces from Proprietary LLM APIsLeading large language model providers now conceal their models' step-by-step reasoning, or chain-of-thought, to protect intellectual property and limit information leakage. Rather than storing these ...
Derrida reposted
Derrida reposted
Derrida reposted
"Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models" arxiv.org/pdf/2511.15304
arxiv.orgarxiv.org
What appears as proof of presence is already a technical mark, iterable in absence, and therefore open to displacement, error, and misuse. GPS shows that location functions like a signature: valid only by its repeatability. #39c3 bsky.app/profile/erle...
You are up to date.